AI Sec Watch

OpenAI is slowing down the release of its new GPT 5.6 model at the request of the US government, offering it first to only a small group of partners instead of a wide public launch. This approach is similar to how Anthropic released its Mythos product, suggesting that AI companies may be coordinating with government oversight when deploying powerful new models.

Amazon Q Developer had a high-severity flaw (CVE-2026-12957, CVSS 8.5) where a malicious repository could run commands and steal a developer's cloud credentials through a configuration file. The bug occurred because Amazon Q automatically launched MCP servers (processes that connect AI assistants to databases and tools) from an untrusted config file without asking the developer for permission first, giving those processes full access to the developer's AWS keys and other sensitive credentials.

Federated learning (FL, a technique where AI models are trained across multiple computers without sending raw data to a central server) is vulnerable to attacks where dishonest participants send corrupted model updates that poison the final model without being detected. This paper introduces AMBER, a framework that adds three layers of verification to check whether clients are trustworthy: confirming data hasn't been tampered with, detecting when clients provide misleading inputs, and verifying that model computations are correct using a trusted execution environment (TEE, a secure area of a computer processor that protects sensitive operations).

This research addresses the challenge of detecting deepfakes (fake videos created by AI) that use multiple types of data like video and audio together, as these deepfake techniques become more advanced. The researchers propose Amber, a system that uses continual learning (where an AI improves over time as it sees new examples) to better remember what real deepfakes look like while adapting to new deepfake methods, inspired by how the immune system selects useful cells.

The Trump administration has asked OpenAI to limit its next model release (GPT 5.6) by vetting initial users before a wider launch, marking the first time a US firm has been told to restrict an AI model before release. OpenAI said each initial partner will be government-approved, and Anthropic is also facing restrictions from Washington.

Amazon Q, an AI coding assistant for VS Code, had a high-severity vulnerability (CVE-2026-12957) that let attackers execute arbitrary code and steal cloud credentials just by having a developer open a malicious repository. The problem was that Amazon Q automatically loaded and ran MCP server configurations (local processes that extend an AI assistant's capabilities) from workspace files without asking the user for permission or checking if the folder was trusted. Since these processes inherited the developer's full environment, attackers could access sensitive credentials like AWS keys and API tokens.

AI agents are autonomous programs that can reason, make decisions, and access multiple systems within enterprises, but the identity governance systems (IAM, or identity and access management, which controls who can access what) were designed for humans and fixed service accounts, not for dynamic autonomous actors. The core problem is that agents inherit all the permissions of the human or service identity they operate on behalf of and can access resources across many systems in a single session without traditional checkpoints, creating a governance gap that existing tools cannot see or control.

OpenAI is previewing GPT-5.6, a new series of AI models including Sol (most powerful), Terra (balanced and cheaper), and Luna (fastest and most affordable). The models include enhanced safety protections against misuse and harmful activities, with Sol featuring improved capabilities in coding, biology, and cybersecurity tasks.

The US has proposed the AI Incident Reporting Act, which would require developers of advanced AI models to report major safety and security incidents to the Commerce Department within seven days of discovering them. The law would cover incidents like attempts to evade human oversight, theft of model weights (the internal parameters that make an AI work), and capabilities that could enable cyberattacks or weapons development, with the Commerce Department required to notify Congress within 48 hours for imminent threats.